What people run on a Mac Studio — and why ours is bigger
Buy a Mac Studio for general local AI and the internet will point you at a ~27 billion parameter model. We run 176 billion — Qwen3.8-Flash-Next — on our own servers. Same family. Not the same league.
By the Blu Solutions team · Updated 21 November 2026 · General information, not advice for your specific situation.
If you asked local-AI people in 2026 what to run on a Mac Studio for general work — writing, coding, thinking, not a science project — the answer that keeps coming back is Qwen3.8-27B.
A 64 GB Studio owner measured that 27B at about 75 tokens per second with thinking left on. Unsloth calls it the strongest model for its size. Mac Studio testers pick it because a dense 70B often fits and then feels like a batch job. For everyday work, the 27B class is the one people actually recommend.
That is a good desktop model. It is not what we run for the firm.
What we run instead
We serve Qwen3.8-Flash-Next: 176 billion total parameters, 256K context, on our servers. You do not install it. You do not need the Studio. See Our AI.
Flash-Next is the checkpoint that does not live comfortably on a Mac. The 27B is the one that does. Ours is the bigger one. On Qwen’s own numbers, it is also the better one.
Qwen’s table — Flash-Next vs the Mac-class 27B
These are Qwen’s published scores, Flash-Next against Qwen3.8-27B. Higher is better. We did not run this bake-off in-house.
| Benchmark | Mac Studio pick (27B) | Blu (Flash-Next, 176B) |
|---|---|---|
| LiveCodeBench v6 | 90.3 | 91.9 |
| SWE-bench Pro | 61.7 | 62.5 |
| SWE-bench Multilingual | 73.8 | 81.0 |
| DeepSWE 1.1 | 42.2 | 58.7 |
| NL2Repo-Bench | 42.3 | 48.1 |
| CoWorkBench | 70.7 | 73.9 |
| JobBench | 33.4 | 55.7 |
Flash-Next wins every row. DeepSWE (42.2 → 58.7) and JobBench (33.4 → 55.7) are not a rounding error. That is why we did not put a 27B under the desk and call it the office model.
For a client
A Mac Studio running 27B is a fine personal setup. An accounting firm’s working papers should not depend on whoever left LM Studio open. We host the 176B model, add our own RAG, and a person still approves the ledger. Data stays in the office.
Sources: Qwen Flash-Next model card (the table), Qwen3.8-27B, Unsloth on 27B, 64 GB Mac Studio measurements.
Where we can help
- Our AI — Flash-Next on our servers, GLM-5.3-Flash for slow reviews
- Accounting + software — Blu OS — firm operating system, not SaaS
- Apply for 1 March — new client intakes